Anthony Ozerov

Simulating turbulence requires accounting for phenomena that occur at scales smaller than the grid scale of the simulation. We combine analytical and deep learning methods to create a better subgrid-scale model and demonstrate better results than using either method alone.”

Turbulence is not just something you feel on a plane, it’s a general phenomenon of chaotic fluid motion which can happen in any liquid or gas. This makes the ability to simulate and characterize turbulence important, from understanding the flow of air over a wing to the long-term behavior of the atmosphere. Any computer simulation of turbulence must occur at some resolution, or “”grid scale.”” But any fluid will have small micro-behaviors at the sub-pixel or “”subgrid”” scale, which turn out to affect behavior at the larger scales. A “”subgrid-scale model”” is therefore needed to account for how these small-scale behaviors interact with the larger scales. We use deep learning to create such a model. Our key advance is that we use an existing gradient-based approximate subgrid-scale model to do most of the heavy lifting, and simply use a neural network to fix the errors in this model. This yields better performance for the model, including in standalone evaluations and when running in the loop of a computer simulation.”